A Wireless Communication Network Resource Allocation Algorithm with Dynamic Adjustment on Demand

Abstract

The present invention discloses a wireless communication network resource allocation algorithm for dynamic adjustment on demand, which specifically includes the following steps: Step 1: Quantitatively describe the characteristics of tasks in the wireless communication network and represent them as input vectors of fixed dimensions #imgabs0# Step 2: Construct a high-reliability and low-latency resource allocation model based on a dynamic neural network with variable widths; Step 3: For the said model, give an optimization objective and a knowledge-driven solution; Step 4: Establish a knowledge base regarding the deployment environment and the optimal inference model; Step 5: Obtain the optimal model from the knowledge base to get the resource allocation scheme. The technical effect of the invention is to divide the network resource allocation problem into two stages: First, determine the optimal width of the decision network according to the task characteristics and the user's computing power; then, input the task characteristics into the optimal decision network to obtain the optimal resource allocation scheme.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless communication network resource allocation, and specifically to a wireless communication network resource allocation algorithm with on-demand dynamic adjustment. Background Art

[0002] In recent years, the fifth-generation (5G) wireless communication network has been commercially available and deployed globally. Although the 5G network can greatly improve network performance and ensure service-level performance, there are still some limitations. For example, the 5G network mainly relies on ground base stations densely deployed in urban areas rather than rural or remote areas, and it is difficult to provide efficient services in some remote and harsh areas.

[0003] With the vigorous development of emerging technologies such as the integrated space-air-ground network, terahertz, and intelligent reflecting surfaces, as well as the deep integration of artificial intelligence and communication technologies, it provides a broad prospect for the research of the sixth-generation (6G) wireless communication network. A major challenge for the 6G network is to provide "on-demand" services. On the one hand, due to the wide coverage, the application of software-defined technology and artificial intelligence technology, a large number of new services with different requirements will emerge. These services and requirements should be processed and satisfied efficiently and economically. On the other hand, the high network dynamics, network architecture heterogeneity, and network resource complexity will jointly lead to unprecedented difficulties in network resource coordination. To solve these problems, the 6G network will comprehensively develop new air interfaces, novel network architectures, and most importantly, advanced artificial intelligence technologies to further improve network performance and provide on-demand services.

[0004] We focus on the resource allocation problem in wireless communication networks. Obviously, in order to obtain higher network performance, it is best to make fast and optimal decisions. However, in most cases, speed (the time required for the allocation decision-making process) and optimality (the gap between the obtained allocation scheme and the optimal scheme) are trade-off metrics. If the trade-off can be adjusted according to specific scenarios and service requirements, it is a reasonable solution. The common method is to formulate the decision problem as an optimization problem and then use an iterative-based method to solve the problem. For example, using an iterative-based method may lead to a large computational complexity. Therefore, for the larger network scale and higher network dynamics of 6G, this method may lead to an overly long algorithm convergence time or even be unable to solve the problem.

[0005] Compared with traditional iterative-based algorithms, the neural network has a lower inference computational cost, so the time required for decision-making is less. Therefore, by reasonably setting and training the network structure, we can also effectively solve the above-mentioned network resource allocation problem using a neural network. However, in the 6G era, different scenarios may lead to different requirements for resource allocation decisions. For example, a base station serving autonomous driving will require a much lower decision latency than one serving Internet of Things monitoring. In addition, the available computing resources on different base stations will also vary. Most neural network models do not have such flexibility that the decision-making speed and optimality can be adjusted according to service requirements and available computing resources. Summary of the Invention

[0006] The purpose of the present invention is to provide a wireless communication network resource allocation algorithm with on-demand dynamic adjustment to solve the problems raised in the above background technology.

[0007] To achieve the above object, the invention provides the following technical solution: a wireless communication network resource allocation algorithm with on-demand dynamic adjustment, specifically including the following steps:

[0008] Step 1: Quantitatively describe the characteristics of the tasks in the wireless communication network and represent them as input vectors of a fixed dimension

[0009] Step 2: Construct a high-reliability and low-latency resource allocation model based on a dynamic neural network with variable width;

[0010] Step 3: For the model, give an optimization objective and a knowledge-driven solution;

[0011] Step 4: Establish a knowledge base regarding the deployment environment and the optimal inference model;

[0012] Step 5: Obtain the optimal model from the knowledge base to get the resource allocation scheme.

[0013] Preferably, in Step 1, the task characteristics are quantified according to three factors: task data volume, task transmission distance, and task importance;

[0014] Among them, the task data volume is in Kb and is represented as s; the task transmission distance is in m and is represented as d; the task importance is the urgency of the task, which is quantified into four different levels as w ∈ [0.8, 0.4, 0.2, 0.1], and the three factors are integrated to describe the task characteristics as

[0015] Preferably, in Step 2, the dynamic neural network with variable width is used as a bandwidth allocation model in the case of multi-task transmission. For different task feature inputs respectively output bandwidth allocation vectors and power allocation vectors

[0016] Preferably, the network resource allocation model consists of an input network, an output network, and several cascaded multi-expert (MoE) layers. Each MoE layer controls the number of expert modules participating in the operation according to different inputs.

[0017] Preferably, the MoE layer consists of an input sub-layer, a gate function network, N parallel expert modules, and an output sub-layer. To make full use of each load module and save computing resources at the same time, the input-output mapping function of the gate function network is set as follows:

[0018] G(X) = RemainK(H(x), k)

[0019] H(X) = X·Wg + Normal·Softplus(X·W noise )

[0020] In the above formula, the first term is the multiplication of the input X and the weight matrix Wg, representing the specificity of different inputs for expert module selection. The second term is the noise term, which is obtained by multiplying the standard positive white noise Normal and the noise coefficient Softplus(X·W noise ) The W noise is the noise matrix. The noise term is used to balance the weights of each expert module. The RemainK function is a sparsification function. According to the given hyperparameter K, the largest K values in the independent variable are retained, and the rest are set to -∞. Since the output of the gate function network is connected to the Softmax layer, only K valid values are finally retained, and the rest are 0. An additional loss term L blance is constructed, and its expression is as follows:

[0021]

[0022] where w blance is a pre-set loss coefficient, CV is the coefficient of variation, representing the degree of dispersion of the weight values of different expert modules. All expert modules have the same structure, which is composed of two cascaded linear layers with activation functions. The first linear layer uses the ReLU function as the activation function, and the second linear layer uses the LogSoftmax function as the activation function.

[0023] Preferably, in step three: Since the computing power of the server is different in different situations, the CPU frequency is used to quantify it, denoted as f. For user i, the task generated by it is denoted as O i , and the total delay for completing a task transmission includes the computing time for the server to allocate bandwidth and the time used in the transmission process. Considering the importance of the task, the delay is weighted and denoted as:

[0024] T i = w i (T com + T tra,i )

[0025] where T i is the total weighted delay of task O i , T com is the server operation delay, and T tra,i is the transmission delay, and w i is the importance of task O i ;

[0026] The operation formulas of T com and T tra,i are respectively:

[0027]

[0028] where n is the number of expert modules used, φ = ||θ1||||θ2|| is the computational complexity of an expert module, ||θ1|| and ||θ2|| are the number of parameters of the first and second linear layers respectively, α is the CPU computational efficiency factor, s i , b i , P i are respectively the size of task O i , the bandwidth used, and the transmission power, d0 is the unit distance, g0 is the channel gain, and σ 2 is the noise power;

[0029] Define the total optimization objective as:

[0030]

[0031] T is the sum of the weighted delays of all tasks; for the dynamic network model, since different allocation models with different hyperparameters K have different operation delays and allocation effects, a knowledge-driven approach is used to obtain the optimal model; traverse all K values in different scenarios, train several dynamic network models with different widths, establish a knowledge base of the deployment environment and the optimal inference model by comparing the model performances, select the one with the lowest total weighted delay as the final allocation model, and use this model to obtain the allocation scheme.

[0032] Preferably, in step four, for the allocation model with a given K value, set the loss function used for training as:

[0033]

[0034] The specific steps to construct the knowledge base of the deployment environment and the optimal inference model are:

[0035] ①Obtain the user task characteristics and server-related information in this deployment environment;

[0036] ②Randomly select a value within the range {1 to N} as the K value of the allocation model, where N is the total number of expert modules in the MoE layer of the expert module, and ensure that the selected K value is different from the previous ones each time;

[0037] ③According to the known task information and loss function, use the backpropagation algorithm to iteratively optimize the parameters θ of the allocation model;

[0038] ④Repeat steps ② to ③ until all models with different K values are completed training, and calculate the corresponding time delay and T according to the allocation schemes output by different inference models respectively, and select the one with the smallest T value as the optimal model and store it in the knowledge base;

[0039] ⑤Repeat the above steps until all considered deployment environments are traversed.

[0040] Preferably, in step ⑤, first select the optimal allocation model from the knowledge base according to the known user and server information, and then use this model to obtain the resource allocation scheme to achieve low-latency multi-task transmission.

[0041] Compared with the prior art, the beneficial effects of the invention are: the proposed on-demand dynamic adjustment of wireless communication network resource allocation algorithm presents a multi-user network resource allocation problem, where tasks generated by each user have different importance. The neural network for decision-making can dynamically adjust the network width according to the characteristics of the tasks and the computing power of the server, and give the optimal solution considering both inference latency and transmission latency; specifically, the bandwidth allocation problem is divided into two stages: first, determine the optimal width of the decision network according to the task characteristics and user computing power; then, input the task characteristics into the optimal decision network to obtain the optimal resource allocation scheme. Brief Description of the Drawings

[0042] Figure 1 is the overall architecture diagram of the resource allocation algorithm;

[0043] Figure 2 is the architecture diagram of the MoE layer;

[0044] Figure 3 is the architecture diagram of all expert modules;

[0045] Figure 4 is the dynamic neural network architecture diagram;

[0046] Figure 5 is the schematic diagram for comparing bandwidth allocation effects;

[0047] Figure 6 is the schematic diagram for comparing power allocation effects. Detailed Embodiment

[0048] Next, in combination with the accompanying drawings in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0049] Please refer to Figures 1-6 , the invention provides a technical solution: a wireless communication network resource allocation algorithm with dynamic adjustment on demand, characterized in that: specifically includes the following steps:

[0050] Step 1: Quantitatively describe the characteristics of the tasks in the wireless communication network and represent them as input vectors of a fixed dimension

[0051] Step 2: Construct a highly reliable and low-latency resource allocation model based on a dynamic neural network with variable width;

[0052] Step 3: For the model, give an optimization objective and a knowledge-driven solution;

[0053] Step 4: Establish a knowledge base about the deployment environment and the optimal inference model;

[0054] Step 5: Obtain the optimal model from the knowledge base to obtain the resource allocation scheme.

[0055] In this embodiment, in Step 1, the task characteristics are quantified according to three factors: task data volume, task transmission distance, and task importance;

[0056] Among them, the task data volume is in Kb and is represented as s; the task transmission distance is in m and is represented as d; the task importance is the urgency of the task, and is quantified into four different levels as w ∈ [0.8, 0.4, 0.2, 0.1]. Integrating the three factors, the task characteristics are described as

[0057] In this embodiment, in Step 2, the dynamic neural network with variable width is used as the bandwidth allocation model in the case of multi-task transmission. As Figure 1 shown in the overall architecture diagram of the bandwidth allocation model, for different task feature inputs respectively output the bandwidth allocation vector and the power allocation vector

[0058] In this embodiment, the bandwidth allocation model consists of an input network, an output network, and several cascaded multi-expert (MoE) layers. Each MoE layer controls the number of expert modules participating in the operation according to different inputs, thereby dynamically adjusting the network width. For example, Figure 2 The following shows the architecture of the MoE layer:

[0059] In this embodiment, the MoE layer is composed of an input sub-layer, a gate function network, N parallel expert modules, and an output sub-layer. Among them, the role of the gate function network is to output the gating values [g1, g2,... g N of the N expert modules according to the variable X transmitted by the input sub-layer and the preset hyperparameter K (K ∈ {1~N}). When the gating value of a certain expert module is 0, the expert module does not participate in the operation, thereby controlling the width of the MoE layer. To make full use of each load module and save computing resources at the same time, the input-output mapping function of the gate function network is set as follows:

[0060] G(X) = RemainK(H(x), k)

[0061] H(X) = X·Wg + Normal·Softplus(X·W noise )

[0062] In the above formula, the first term is the multiplication of the input X and the weight matrix Wg, representing the specificity of different inputs for the selection of expert modules. The second term is the noise term, which is obtained by multiplying the standard normal white noise Normal and the noise coefficient Softplus(X·W noise ) and W noise is the noise matrix. The noise term is used to balance the weights of each expert module. The RemainK function is a sparsification function. According to the given hyperparameter K, the largest K values in the independent variable are retained, and the remaining values are set to -∞. Since the output of the gate function network is connected to the Softmax layer, only K valid values are finally retained, and the remaining values are all 0;

[0063] To make full use of each expert module and avoid the situation where only a very small number of expert modules have large weights while the weights of the remaining majority of modules are low or 0, an additional loss term L blance is constructed, and its expression is as follows:

[0064]

[0065] Among them, w balance is the preset loss coefficient, and CV is the coefficient of variation, representing the degree of dispersion of the weight values of different expert modules.

[0066] For example, Figure 1As shown in the overall architecture diagram of the bandwidth allocation model, the main part of the allocation model is the MoE layer with a gating unit network. To enhance the learning ability of the model and obtain better allocation performance, the structures of all expert modules are set as follows Figure 3 As shown: The expert module is composed of two cascaded linear layers with activation functions. The ReLU function is used as the activation function for the first linear layer, and the LogSoftmax function is used as the activation function for the second linear layer to reduce the difficulty of network training and improve the convergence speed.

[0067] In this embodiment, step three: Considering the scenario of multiple users and a single server, each user generates tasks locally and transmits them to the server for processing; the total bandwidth and total power occupied by all users are fixed. The server uses the allocation model to allocate resources according to the task information of different users; due to the diversity of server devices and deployment environments, the computing capabilities of the server are different in different scenarios. The CPU frequency is used to quantify it, denoted as f. For user i, the task generated by it is denoted as O i The total delay for completing one task transmission includes the computing time for the server to allocate bandwidth and the time used in the transmission process. Considering the importance of the task, the delay is weighted and expressed as:

[0068] T i = w i (T com + T tra,i )

[0069] Among them, T i is the total weighted delay of task O i , T com is the server computing delay, T tra,i is the transmission delay, w i is the importance of task O i ;

[0070] The calculation formulas for T com and T tra,i are respectively:

[0071]

[0072] Among them, n is the number of expert modules used, φ = ||θ1||||θ2|| is the computational complexity of an expert module, ||θ1|| and ||θ2|| are the number of parameters of the first and second linear layers respectively, α is the CPU computing efficiency factor, s i , b i , P i are respectively the size, the bandwidth used, and the transmission power of task O i , d0 is the unit distance, g0 is the channel gain, and σ 2 is the noise power;

[0073] Since there are multiple users for task transmission, the total optimization objective is defined as:

[0074]

[0075] T is the weighted delay sum of all tasks; for the dynamic network model, since different allocation models with different hyperparameters K have different operation delays and allocation effects, a knowledge-driven approach is used to obtain the optimal model; the K value is traversed and trained in different scenarios to obtain several dynamic network models with different widths, and by comparing the model performances, a knowledge base of the deployment environment and the optimal inference model is established, and the one with the lowest total weighted delay is selected as the final allocation model, and the allocation scheme is obtained using this model.

[0076] In this embodiment, in step four, for the allocation model with a given K value, the loss function used for training is set as:

[0077]

[0078] The specific steps for constructing the knowledge base of the deployment environment and the optimal inference model are as follows:

[0079] ① Obtain the user task characteristics and server-related information in this deployment environment;

[0080] ② Randomly select a value within the range {1 to N} as the K value of the allocation model, where N is the total number of expert modules in the MoE layer of the expert module, and ensure that each selected K value is different from the previous ones;

[0081] ③ According to the known task information and the loss function, use the backpropagation algorithm to iteratively optimize the parameters θ of the allocation model;

[0082] ④ Repeat steps ② to ③ until all models with different K values are trained, and calculate the corresponding delay sum T according to the allocation schemes output by different inference models, and select the one with the smallest T value as the optimal model and store it in the knowledge base;

[0083] ⑤ Repeat the above steps until all considered deployment environments are traversed.

[0084] In this embodiment, in step five, first select the optimal allocation model from the knowledge base according to the known user and server information, and then use this model to obtain the resource allocation scheme to achieve low-delay multi-task transmission, as Figure 4 shown.

[0085] Technical effect: The on-demand dynamic adjustment wireless communication network resource allocation algorithm proposes a multi-user network resource allocation problem, where tasks generated by each user have different importance. The neural network for decision-making can dynamically adjust the network width according to the characteristics of tasks and the computing power of the server, and give the optimal solution considering both inference latency and transmission latency. Specifically, the bandwidth allocation problem is divided into two stages: First, determine the optimal width of the decision network according to task characteristics and user computing power; then, input the task characteristics into the optimal decision network to obtain the optimal bandwidth allocation scheme. Working principle: In 6G wireless communication networks, on-demand services are a key but challenging issue because emerging service requirements are significantly diverse and network resources become increasingly dynamic. By studying the on-demand wireless network resource allocation problem and focusing on the computing latency problem during the resource allocation process. Specifically, the decision latency is considered in the optimization problem, and then the on-demand dynamic adjustment wireless communication network resource allocation algorithm is proposed, where the model computing complexity can be adjusted according to service requirements. Further, a knowledge base is constructed to represent the relationship between service requirements, available computing resources, and resource allocation performance. By leveraging this knowledge, the appropriate width of the dynamic neural network is selected to further optimize the resource allocation effect. As Figure 5 and Figure 6 shown, Static net is a traditional static neural network without a gating network; dynn is a dynamic neural network with a fixed K value; knoelwdge-dynn is the algorithm of this application. The simulation results show that this algorithm is significantly better than the traditional static neural network, significantly reducing the latency (including inference latency and transmission latency), and having higher flexibility in providing on-demand services.

[0086] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention also intends to include these changes and modifications.

Claims

1. A wireless communication network resource allocation algorithm with dynamic adjustment on demand, characterized in that: Specifically, it includes the following steps: Step 1: Quantitatively describe the characteristics of the tasks of the wireless communication network and represent them as input vectors of fixed dimensions Step 2: Construct a highly reliable and low-latency resource allocation model based on a dynamic neural network with variable width; Step 3: For the said model, give an optimization objective and a knowledge-driven solution; Step 4: Establish a knowledge base regarding the deployment environment and the optimal inference model; Step 5: Obtain the optimal model from the knowledge base to get the resource allocation scheme; In step three: Since the computing power of the server varies in different scenarios, the CPU frequency is used to quantify it, denoted as f. For user i, the task generated by the user is denoted as O i , the total delay for completing a task transmission includes the computing time for the server to allocate bandwidth and the time used in the transmission process. Considering the importance of the task, the delay is weighted and denoted as: T i = w i (T com + T tra,i ) Among them, T i is the total weighted delay of task O i , T com is the server computing delay, T tra,i is the transmission delay, w i is the importance of task O i ; T com and T tra,i The operation formulas are respectively as follows: where n is the number of expert modules used, φ = ||θ1|| ||θ2|| is the computational complexity of one expert module, ||θ1|| and ||θ2|| are the number of parameters of the first and second linear layers respectively, α is the CPU computational efficiency factor, s i , b i , P i are the size, bandwidth used, and transmission power of task O i respectively, d0 is the unit distance, g0 is the channel gain, and σ 2 is the noise power; Define the total optimization objective as: T is the weighted delay sum of all tasks; for the dynamic network model, since the allocation models with different hyperparameters K have different operation delays and allocation effects, a knowledge-driven approach is used to obtain the optimal model; traverse all K values in different situations, train several dynamic network models with different widths, establish a knowledge base of the deployment environment and the optimal inference model by comparing the performances of each model, select the one with the lowest total weighted delay as the final allocation model, and use this model to get the allocation scheme; In Step 4, for the allocation model with a given K value, set the loss function used for training as: The specific steps to construct the knowledge base regarding the deployment environment and the optimal inference model are: ① Obtain the user task characteristics and server-related information in this deployment environment; ② Randomly take values in the range {1~N} as the K value of the allocation model, where N is the total number of expert modules in the Mixture of Experts (MoE) layer of the expert module, and ensure that each selected K value is different from the previous ones; ③ According to the known task information and the loss function, use the backpropagation algorithm to iteratively optimize the parameters θ of the allocation model; ④ Repeat steps ②~③ until all models with different K values are completed training, and calculate the corresponding delay sum T respectively according to the allocation schemes output by different inference models, and select the one with the smallest T value as the optimal model and store it in the knowledge base; ⑤ Repeat the above steps until all considered deployment environments are traversed.

2. The on-demand dynamic adjustment wireless communication network resource allocation algorithm according to claim 1, characterized in that: In Step 1, the task characteristics are quantified according to three factors: task data volume, task transmission distance, and task importance; Among them, the task data volume is in the unit of Kb and is denoted as s; the task transmission distance is in the unit of m and is denoted as d; the task importance is the urgency of the task, which is quantified into four different levels as w ∈ [0.8, 0.4, 0.2, 0.1]. Integrating the three factors, the task characteristics are described as 3. A wireless communication network resource allocation algorithm for dynamic adjustment on demand as claimed in claim 1, characterized in that: In step two, the dynamic neural network with variable width is used as the bandwidth allocation model in the case of multi-task transmission, and for different task feature inputs respectively output the bandwidth allocation vector and the power allocation vector 4. The on-demand dynamic adjustment wireless communication network resource allocation algorithm according to claim 3, wherein: The bandwidth allocation model consists of an input network, an output network, and several cascaded Mixture of Experts (MoE) layers. Each MoE layer controls the number of expert modules participating in the operation according to different inputs and the hyperparameter K.

5. The on-demand dynamic adjustment wireless communication network resource allocation algorithm according to claim 4, characterized in that: The MoE layer consists of an input sublayer, a gate function network, N parallel expert modules, and an output sublayer; in order to make full use of each load module and save computing resources at the same time, set the input-output mapping function of the gate function network as follows: G(X) = RemainK(H(x),k) H(X) = X·Wg + Normal·Softplus(X·W noise ) In the above formula, the first term is the multiplication of the input X and the weight matrix Wg, representing the specificity of different inputs for the selection of the expert module. The second term is the noise term, which is obtained by multiplying the standard positive white noise Normal by the noise coefficient Softplus(X·W noise ), where W noise is the noise matrix. The noise term is used to balance the weights of each expert module. The RemainK function is a sparsification function. According to the given hyperparameter K, it retains the largest K values in the independent variable and sets the remaining values to -∞. Since the output of the gating function network is connected to the Softmax layer, finally only K valid values (non-zero) are retained, and the remaining values are all 0. An additional loss term L blance is constructed, and its expression is as follows: where w blance is a pre-set loss coefficient, and CV is the coefficient of variation, representing the degree of dispersion of the weight values of different expert modules; all expert modules have the same structure, which is composed of two cascaded linear layers with activation functions. The ReLU function is used as the activation function for the first linear layer, and the LogSoftmax function is used as the activation function for the second linear layer.

6. The on-demand dynamic adjustment wireless communication network resource allocation algorithm according to claim 1, characterized in that: In Step 5, first select the optimal allocation model from the knowledge base according to the known user and server information, and then use this model to get the resource allocation scheme to achieve low-latency multi-task transmission.

Images